by tom's HARDWARE January 13 2017 Click here for a complementary article written circa 1995, discussing how a memory controller has to be "re-programmed" for the proper RAM configuration in a PC. I am wondering what is a memory bank and what exactly it does? Does it effect ram in any way? A memory bank is a functionally independent memory device that is constructed in parallel with other banks in order to increase the throughput of the memory architecture. Many types of RAM and even some non-volatile storage use multiple memory banks. For the purposes of keeping things brief and not overly technical I'll detail how memory banks work in DDR3 (launched in 2007) and DDR4 (launched in 2014) SDRAM. Go to DDR5 review (launched in 2020) but not yet in mainstream production. Click here for GDDR6, the latest Graphics chip by Samsung launched in 2016. It was the successor to GDDR5 launched in 2007, based on DDR3. All DDR3 chips contain eight fully independent memory banks tied to a shared IO interface. A 4 gigabit DDR3 die (the largest capacity produced) is actually constructed from eight 512 megabit memory banks. During operation, each rank has one row open from all of the rows in its own memory – a memory bank cannot open a row in another bank's memory. When a command is sent to the selected rank, the Bank Address is used to determine which memory bank the command is to be sent to. The chip's internal decoding circuitry takes care of the rest. Since the IO bus is shared, care must be taken to avoid sending commands to banks such that the results may overlap. Consider the following: A CPU contains a single dual-channel DDR3 SDRAM controller. Each memory channel is populated by two DIMMs. Each DIMM contains two ranks. Each rank is constructed from 4 gigabit DDR3 SDRAM chips. Now, lets do some math. <2 DIMMs per channel> * <2 ranks per DIMM> = 4 ranks per channel <4 ranks per channel> * <8 banks per rank> = 32 banks per channel. Since each bank has its own open row, more banks is better! <8 kilobits per row> * <65,536 rows per bank> = 512 megabits per bank <512 megabits per bank> * <8 banks per die> = 4 gigabits per die <64 IO bits per rank> / <8 IO bits per die (in this example)> = 8 chips per rank <8 chips per rank> * <1 die per chip> * <4 gigabits per die> = 32 gigabits per rank <32 gigabits per rank> / <8 bits per byte> = 4 gigabytes per rank <4 gigabytes per rank> * <2 ranks per DIMM> = 8 gigabytes per DIMM <8 gigabytes per DIMM> * <2 DIMMs per channel> = 16 gigabytes per channel <16 gigabytes per channel> * <2 channels> = 32 gigabytes installed Since each bank has one row out of 65,536 open, the memory controller must know the precise state of each bank under its command. The greater the number of banks installed, the greater the likelihood that the memory controller will be able to reorder and schedule a command without having to open a new row on a bank, which is a time consuming task. DDR4 works slightly differently. Whereas DDR3 has eight fully independent banks, DDR4 has four fully independent bank groups, with each bank group having four semi-independent banks within it. The principle and mathematics behind it are the same but DDR4 has additional timing constraints not present in DDR3.
DDR5 is the next generation of RAM that has been the subject of speculation for the past couple of years. The release of its standard specification was originally scheduled for 2018 but was pushed back repeatedly, its official release date finally being 14th July 2020. The first DDR5 chip actually revealed was by SK Hynix on 6th October 2020, but this was targeted for specialist “Big Data, Artificial Intelligence, and Machine Learning” work rather than gaming. It is likely the first examples for servers will be available earlier in 2021, with widespread availability for desktop PCs being in the latter half of 2021. Typically with RAM, more so than other computer components like CPUs and GPUs, there is a significant lag time before other computer hardware catches up to actually take full advantage of the technology. For DDR5, as was the case with previous generations of RAM, the latest motherboards and processors (even the recent AMD 5000 series CPUs) are not DDR5 compatible, though it is expected that Intel 12th Gen Alder Lake, scheduled for late 2021 will be. There is as yet no official word on when we can expect AMD CPUs to support DDR5, though a recent leak reported by GamingNexus on AMD’s roadmap for CPU development shows Zen 4 CPUs, scheduled for a 2022 release with DDR5 compatibility. This being the case, DDR5 is not even an option at present. The question is, when it is released, will it be worth the upgrade? Let’s take a look at what we do know. The Specifications As we can see below, DDR5 promises improvements on DDR4 in the key areas which affect gaming and workstation users i.e.: memory capacity, speed, and also power efficiency. As we will discuss though, much of these improvements will not have much of an impact for desktop computers for some time. DDR5 Power Consumption Although lower power consumption for a more powerful chip might seem counterintuitive, the shrinking transistor sizes for the new generation mean increases in both, as was the case with previous DDR generations. The 0.1V (8.4%) reduction might not seem much, and likely won’t make a great deal of difference for PC users, but should mean noticeable increases in battery life for mobile phones when the technology is rolled out there. Interestingly DDR5 chips depart from previous generations in that they will regulate their own voltage, as opposed to this being a feature controlled from the motherboard. This may (we hope) reduce the costs of motherboards in the future. DDR5 Compatibility It is important to note that, as is the case with DDR3 and DDR4, the pin layout for DDR5 will be different from preceding generations. This means DDR4 motherboards and DDR5 memory won’t be cross-compatible. DDR5 Error Correcting Code (ECC) Memory Error checking is a feature that is most important for servers, which must remain online at all times. DDR5 promises improved error checking over DDR4, and to transfer full control of this feature ‘on die’ (i.e. on the RAM chips themselves). Removing the memory controller overseeing this function from the CPU onto the RAM itself should bring gains in processing power. DDR5 Data Rate & Clock Frequency DDR RAM (Double Data Rate Random Access Memory) is so called because it carries double the information on each clock cycle than older versions of RAM could, by utilizing the “falling” part of the digital signal, as well as the “upward” part, to carry information. As a result, although the effective transfer rate of DDR3-2000 for instance would be 2000Gbps, its actual clock frequency would be 1000MHz. The “stock” data rate for DDR4 is 1600-3200Gbps, and for DDR5 is 4800-6400Gbps (i.e. double), though in reality manufacturers up this frequency beyond this substantially. There is DDR4 memory available on the market for 5100Gbps, though this will cost you around $900 and is a virtually pointless expense for gaming. Overclocking RAM in the BIOS can also bring additional gains. Ultimately DDR5 offers the potential for “off-spec” clock frequencies to be pushed even further, and more room for overclocking, though how long it will take for these to appear, and how useful they will actually be for gamers at least, is not clear. With RAM specifications you typically get diminishing returns when you go beyond what the CPU and other hardware need to operate at maximum efficiency. We know for instance that the new AMD Ryzen 5000 series processors operate best with RAM at around 4000GBps or 2000MHz. DDR5 Memory Timings and Latency It is expected that DDR5 will have about the same latency as DDR4, although technological improvements enabling the memory to be accessed, even whilst some of its memory banks are being refreshed, may give some improvements over its predecessor, at least after the technology is fine-tuned. Ultimately we do not have access to any information on memory timings for DDR5 as of yet. This key metric, written on the front of a RAM chip as four numbers separated by hyphens, indicates the number of clock cycles it takes for the memory to retrieve data. Reading from left to right these four numbers are the CAS, tRCD, tRAP, and tRAS; all of which represent the number of cycles it takes for the RAM to perform a specific operation. For instance, some DDR3-2000 RAM would have an actual clock frequency of 1000MHz (2000GBps effective transfer rate/2), meaning a clock cycle of 1 second/1000 = 1ns (nanosecond); if the CAS number was 6, this would indicate this procedure would take 6ns Historically, the earliest DDR4 chips, despite having higher clock speeds and memory size than their DDR3 predecessors, actually proved slower on release due to having substantially higher timings. It took manufacturers some time to refine the technology to a point it outperformed the preceding generation. Only time will tell whether the same proves true for DDR5. DDR5 Memory Capacity We know that DDR5 has the potential for up to 4x the memory of DDR4: up to 64GB per chip for UDIMM memory – i.e. that used for desktop PCs. There is the potential to stack up 8x of these chips when the motherboard technology becomes available (likely 2022 at the earliest), meaning a system RAM of 512GB, which is an absurd amount of memory. Bear in mind that until recently your low-medium spec gaming setup only really needed 8GB of RAM, though we now recommend 16GB as standard. When testing the new Ryzen 5000 series processors we found that the jump from 16GB to 32GB did make a substantial difference to performance. AMD’s architecture, specifically the Infinity fabric, requires sufficient memory capacity in order to get the most out of it. For video editors, 32GB of RAM would be the standard, with an upgrade to 64GB if you were handling 4K footage. Still, taking your RAM up to 64GB can be easily accomplished using 4x sticks of DDR4. Even if we imagine a situation where 64GB still isn’t sufficient for a desktop user over the next year, topping up the RAM with additional 16GB DDR4 sticks would likely be the most viable option, rather than having to transition to DDR5. How much does RAM Affect Performance? Historically RAM was quite low down the priority list when it comes to components for a new gaming build. Although a minimum amount of RAM capacity is absolutely necessary, particularly if you are looking to build a PC that can be used for video editing and other workstation tasks as mentioned, the uplift more expensive RAM brings is often quite understated compared to say spending more money on a superior GPU. Nevertheless, in some instances, better RAM can mean noticeable increases in Frames Per Second (FPS) for instance, when playing games. Increases from RAM can be particularly notable when playing games on 1080p resolution. The more you up the resolution of the game above 1080p, the more it becomes GPU bound instead of CPU bound. 1080p gaming is, therefore, more reliant on CPU performance and improving both the memory speed and size improves the performance of the CPU; you do get diminishing returns however the higher the resolution goes. Similarly, certain game titles are more dependent on CPU performance, such as Microsoft Flight Simulator and Kerbal Space Programme, where complex physics simulations are required, or in large persistent world titles. FPS performance on these games can be improved noticeably through upgrading your RAM, as this removes bottlenecks on your CPU’s performance. By the same logic “onboard graphics”, such as AMD’s AMU, utilize onboard RAM as VRAM, so upgraded RAM may benefit your average laptop users more. As mentioned, AMD’s Zen 2 CPUs were more dependent on RAM than their Intel rivals. 16GB was generally required to get the most out of the higher end of the Ryzen 3000 series. Zen 3 architecture, according to AMD themselves, has a sweet spot of 4000GBps RAM speed in order to maximize CPU performance. Our own testing of the new AMD Zen 3 processors has found that the jump from 16GB of RAM to 32GB made a huge difference in terms of FPS. This clearly shows the performance bottlenecking that can occur if RAM is not sufficient, however, it is also the case that anything above 4000GBps and 32GB of RAM (all of which can be met with DDR4) would not have much positive difference at all to CPU performance, unless of course improved timings see a reduction in latency. This illustrates the fact that the potential benefits of DDR5 will not be felt until CPU, motherboard, and GPU technology develops to a point where this potential can be unlocked. In Summary Whilst the benefits that DDR5 offers are exciting, the reality is we will not see PCs begin to utilize these benefits until this time next year, and even then, it is unlikely that we will see CPUs able to fully unlock the potential of DDR5 until beyond 2022. DDR4 was released back in 2014 and we still see some gaming machines using DDR3 to this date, though it is increasingly outdated now. At the point DDR4 was released we did not see much difference from its predecessor in performance, and this was the case for several years. Similarly, DDR5 is unlikely to be widespread in most computers until late 2021, and we can expect DDR4 will still be a big chunk of the market until late 2022. If you end up buying a DDR5 compatible Intel or AMD CPU when they come out in 2021 then DDR5 RAM might end up being worth it, though it is impossible to say at this stage what the performance difference might be, if any. If you are intending to hold on to your current CPU, or indeed, are buying one of the new Ryzen 5000 series, then upgrading to DDR5 won’t be worth bothering with until you next decide to upgrade your computer with DDR5 compatible components. Ultimately DDR5 will raise the ceiling on the potential performance CPUs and gaming PCs, in general, can reach, but until these other components develop to a level that can actually meet this, an upgrade to DDR5 will likely be pointless. Although better RAM can make a difference to performance, there is still a threshold beyond which additional increases mean negligible improvements, and for the next year or so at least, DDR4 looks to be sufficient to meet this threshold.
** End of Page